feat: resilient retry sandbox for profile aggregation and write-back jobs

[ibloud]

Summary

Add a retry sandbox for asynchronous profile pipeline steps (aggregation and external write-back) to isolate transient failures and guarantee eventual consistency.

Why

Cloud/network/API failures should not drop profile updates or corrupt state.

Scope

• Introduce retry policy with bounded exponential backoff + jitter.
• Classify retryable vs terminal errors.
• Persist retry metadata (attempt count, last error, next retry time).
• Add dead-letter path after max attempts.
• Ensure idempotent reprocessing of the same match/profile event.
• Add observability counters:
  • success, retry, terminal-fail, dead-letter, duplicate-suppressed

Data/Schema

• Define retry sandbox collection/document shape.
• Record correlation IDs to trace event lineage end-to-end.

Acceptance Criteria

• Transient failures recover automatically within policy bounds.
• Terminal failures are quarantined with actionable diagnostics.
• Replays do not double-apply trait/profile mutations.
• Metrics/logs allow operators to inspect retry health.